A Bayesian model for human directional localization of broadband and static sound sources

Abstract

Humans estimate sound-source directions by combining prior beliefs with sensory evidence. Prior beliefs represent statistical knowledge about the environment while sensory evidence is acquired from auditory features such as interaural disparities and monaural spectral shapes. Models of directional sound localization often impose constraints on the contribution of these features to either the horizontal or vertical dimension. Instead, we propose a Bayesian model that more flexibly incorporates each feature according to its spatial precision and integrates prior beliefs in the inference process. We applied the model to directional localization of a single, broadband, stationary sound source presented to a static human listener in an anechoic environment. We simplified interaural features to be broadband and compared two model variants, each considering a different type of monaural spectral features: magnitude profiles and gradient profiles. Both model variants were fitted to the baseline performance of five listeners and evaluated on the effects of localizing with non-individual head-related transfer functions (HRTFs) and sounds with rippled spectrum. The model variant with spectral gradient profiles outperformed other localization models. This model variant appears particularly useful for the evaluation of HRTFs and may serve as a basis for future extensions towards modeling dynamic listening conditions.